Between 2015 and 2050, half of the net increase in the world's urban populationis expected to take place in Sub-Saharan Africa (SSA), driving drastic landcover changes and challenging the spatial organization of human societies.Understanding past and present dynamics of this urbanization process is criticalto achieve a sustainable pattern of urban development, yet is limited by thelack of accurate and multi-temporal spatial data on urban expansion. Since the2000s, the rise of satellite-based Earth Observation (EO) enabled the productionof several global urban maps, thereby mitigating the issue of data scarcity. ButSSA is still characterized by lower accuracies in satellite-based maps becauseof various issues, such as: a lower satellite imagery availability, a lack ofreference datasets, and a high heterogeneity across the urban areas of theregion.In this thesis, I propose to leverage open-access satellite catalogs along withvolunteered geographic information to improve large-scaled and automated mappingof the built environment in SSA. The proposed approach makes use ofOpenStreetMap to support model training and calibration, thereby bypassing theneed for reference datasets or manual digitization campaigns. This method wasassessed in 10 urban areas of SSA, reaching classification performances similarto manual approaches.Moreover, the combined use of multispectral and synthetic-aperture radar (SAR)imagery was explored. In 11 out of 12 case studies in SSA, multi-sensorclassification schemes outperformed single-sensor approaches. More specifically,multi-sensor classification dramatically increased built-up detection rates inarid and semi-arid regions---where bare soil and buildings may share a similarspectral signature.These findings were implemented to map the built environment of 46 urban areasat five different dates from 1995 to 2015, with an average F1-score of 0.93. Thestatistical interpretation of the produced dataset revealed the highheterogeneity that characterizes urban areas in SSA, and confirmed that thespatial patterns of urbanization highly depends on demographic and economicfactors. Overall, the present thesis provides promising insights for large-scaled andautomated mapping of the built environment in data-scarce regions. Severalissues are still affecting the mapping accuracies, such as: multi-temporalinconsistencies caused by the use of imagery from 7 different sensors, lowavailability of historical imagery in SSA, or missing data in OpenStreetMap.Still, with the growing availability of open-access EO catalogs and theincreasingly completeness of OpenStreetMap, the proposed approach is expected tobecome even more relevant in the near future. / Doctorat en Sciences agronomiques et ingénierie biologique / info:eu-repo/semantics/nonPublished
Identifer | oai:union.ndltd.org:ulb.ac.be/oai:dipot.ulb.ac.be:2013/305158 |
Date | 14 May 2020 |
Creators | Forget, Yann |
Contributors | Gilbert, Marius, Coheur, Pierre-François, Lennert, Moritz, Van Rompaey, Anton, Vandevoorde, Tim T. |
Publisher | Universite Libre de Bruxelles, Université libre de Bruxelles, Faculté des Sciences – Ecole Interfacultaire des Bioingénieurs, Bruxelles |
Source Sets | Université libre de Bruxelles |
Language | English |
Detected Language | English |
Type | info:eu-repo/semantics/doctoralThesis, info:ulb-repo/semantics/doctoralThesis, info:ulb-repo/semantics/openurl/vlink-dissertation |
Format | 1 v. (125 p.), 3 full-text file(s): application/pdf | application/pdf | application/pdf |
Rights | 3 full-text file(s): info:eu-repo/semantics/openAccess | info:eu-repo/semantics/openAccess | info:eu-repo/semantics/closedAccess |
Page generated in 0.0028 seconds